Assess the AI risk category

Detailed guide for classifying AI systems according to AI Act risk categories.

Category Overview

Prohibited Practices (Article 5)

For a practice prohibited by Article 5, assess the precise conduct prohibited, conditions, exceptions and applicable date of the relevant provision.

List of Prohibited Practices

This is an orientation to Article 5, not a ban on every system with the named feature. Assess the purpose, conditions and exceptions of the specific provision.

Practice	Decisive conditions
Manipulation and deception	Material distortion of informed decision-making causing or reasonably likely to cause significant harm (point a)
Exploiting vulnerabilities	Age, disability or a specific social/economic situation; material distortion of behaviour and significant harm (point b)
Social scoring	Detrimental treatment in unrelated contexts, or unjustified/disproportionate treatment (point c)
Predicting an individual’s criminal conduct	Based solely on profiling or personality traits; distinguish permitted support for human assessment based on objective facts (point d)
Facial recognition databases	Untargeted scraping of internet/CCTV facial images to create or expand a database (point e)
Emotion recognition at work or in education	Exception for medical or safety reasons (point f)
Sensitive biometric categorisation	Subject to the scope and exceptions in point g; not all biometric categorisation
Real-time biometric identification	Remote identification in publicly accessible spaces for law enforcement; narrow statutory exceptions and safeguards (point h)
New prohibitions concerning non-consensual intimate material and child sexual abuse material	Points ba, bb and paragraphs 1a–1b apply from 2 December 2026

Self-assessment Checklist

We assessed subliminal, purposefully manipulative or deceptive techniques against all conditions of point (a), including impairment of informed decision-making and actual or reasonably likely significant harm.
We assessed exploitation of vulnerabilities under point (b), including the specified grounds, distortion of behaviour and actual or reasonably likely significant harm.
We assessed social scoring and the resulting detrimental treatment against point (c).
We assessed individual criminal-risk prediction against point (d), including its exception for support of human assessment.
We assessed untargeted scraping to create or expand facial recognition databases under point (e).
We assessed emotion recognition at work or in education under point (f), including the medical/safety exception.
We assessed sensitive biometric categorisation against the scope and exceptions of point (g).
We assessed real-time remote biometric identification in publicly accessible spaces for law enforcement against point (h) and its statutory safeguards.
For 2 December 2026, we assessed points (ba)/(bb) on intimate or sexually explicit material and child sexual abuse material together with paragraphs 1a/1b; system capability alone does not settle all conditions.

A checked item records a completed assessment, not an automatic conclusion that every system with the feature is prohibited. If the actual conduct meets an applicable prohibition, stop that conduct and consult Legal.

High-Risk AI (Article 6, Annexes I and III)

Annex III --- High-Risk Areas

Area	Examples	Article
1. Biometrics	Remote identification (excluding solely identity verification), sensitive/protected categorisation and emotion recognition, where legally permitted	III.1
2. Critical infrastructure	Safety components in critical digital infrastructure, road traffic, water, gas, heating and electricity supply	III.2
3. Education	Admissions, grading, proctoring	III.3
4. Employment	Recruitment, HR decisions, performance	III.4
5. Essential services	Specified public assistance by public authorities or on their behalf; natural-person creditworthiness (excluding financial fraud detection), life/health insurance risk/pricing, emergency calls and healthcare triage	III.5
6. Law enforcement	Profiling, polygraph, risk assessment	III.6
7. Migration	Visa decisions, border control	III.7
8. Justice and democracy	Assistance to courts or on their behalf in applying facts and law, analogous ADR; intended influence on elections/referendums or voting, excluding non-exposed administrative/logistical campaign tools	III.8

Detailed High-Risk AI Examples

High-risk status depends on the specific intended purpose, not a sector label. Article 6(1) requires the Annex I product/safety-component route and the relevant mandatory third-party assessment; assess paragraphs 1a–1c as well. Annex I Section B has the specific scope in Article 2(2).

For Annex III, Article 6(3) permits an exception where there is no significant risk of harm to health, safety or fundamental rights, including through material influence on decision-making, and at least one specified condition applies: a narrow procedural task; improving a completed human activity; detecting decision-making patterns/deviations without replacing or influencing a completed human assessment without proper human review; or a preparatory task for an Annex III assessment. An Annex III system profiling natural persons is always high-risk. A provider relying on this exception documents the assessment before placing on the market or putting into service and registers under Articles 6(4) and 49(2).

Creditworthiness is point 5(b), excluding financial fraud detection; life/health insurance risk assessment and pricing is point 5(c); emergency healthcare triage is point 5(d). A medical device may also follow Article 6(1); product and Annex III routes can overlap, with Article 43(3) governing the conformity route for Annex I Section A products.

The following requirements primarily concern high-risk providers. Deployers follow Article 26; FRIA applies only to the deployers and uses specified in Article 27. DPIA is required where GDPR Article 35 triggers it, not automatically for every AI system. The Article 113 deferral to 2 December 2027 (Annex III) or 2 August 2028 (Article 6(1)) concerns Chapter III Sections 1–3, except Article 6(5). Assess other provisions, including Article 43, and Article 111 transitions separately.

High-Risk Obligations

Obligation	Description	Deadline
Risk management	Lifecycle risk management system (Art. 9)	Ongoing; testing before placing on the market or putting into service
Data Governance	Data quality, bias testing, documentation	Ongoing
Technical documentation	Article 11 and Annex IV documentation; a model card alone is insufficient	Before placing on the market or putting into service, then keep updated
Record Keeping	Logs under the actor’s control: appropriate period, at least 6 months unless other law provides otherwise (Arts. 19, 26(6))	Ongoing
Transparency to deployers	Sufficient interpretability and instructions for use (Art. 13)	Before placing on the market or putting into service
Human oversight	Measures proportionate to risks, autonomy and context (Art. 14)	System design and use
Accuracy & Robustness	Testing, monitoring	Ongoing
Conformity assessment	Applicable Article 43 route: internal control, notified body or sectoral procedure	Before placing on the market or putting into service

Limited Risk / Transparency (Article 50)

Systems Requiring Transparency

System	Obligation
Chatbots	Providers ensure natural persons know they interact with AI unless obvious in context; Article 50(1) includes a law-enforcement exception
Generative AI	Providers: machine-readable marking of synthetic outputs under Art. 50(2), with exceptions and the Art. 111(4) transition. Deployers: disclosure for deepfakes and specified public-interest text under Art. 50(4), not every AI-assisted text
Deepfakes	Deployers disclose generated/manipulated content; Art. 50(4) has a law-enforcement exception and limited disclosure for manifestly artistic, creative, satirical, fictional or analogous works or performances
Emotion recognition	Deployers inform exposed persons and comply with applicable personal-data law; subject to Art. 50(3)’s law-enforcement exception
Biometric categorisation	Deployers inform exposed persons and comply with applicable personal-data law; subject to Art. 50(3)’s law-enforcement exception

Article 50 duties can apply alongside high-risk duties. Assess each statutory exception, including the conditions for editorially reviewed public-interest text under paragraph 4. Information under paragraphs 1–4 must be clear and distinguishable by the first interaction or exposure and meet applicable accessibility requirements (paragraphs 5–6). The labels below are illustrations; visible wording alone does not fulfil machine-readable marking under paragraph 2.

Implementing Transparency

Examples of correct labelling:

CHATBOT:
"You are communicating with an AI assistant. To speak
with a human, type 'operator'."

GENERATED CONTENT:
"This text was created using AI."
"AI-generated image"

DEEPFAKE:
"This content was created or modified using AI."

Minimal Risk

Possible examples after assessing the intended purpose

The table contains possible examples, not automatic legal categories. Assess the specific intended purpose against Articles 5 and 6, and assess Article 50 separately, before using the practical label “minimal risk”. Internal analytics and automation alone do not determine the applicable duties.

System	Examples
Spam filters	Email spam detection
Internal analytics	Business intelligence, dashboards
Recommendation	Product recommendations (no impact on rights)
Search	Internal search, document search
Automation	Workflow automation, scheduling

Article 4 literacy and Article 5 prohibitions still apply, as do relevant GDPR and other legal duties. “Minimal risk” is a practical description, not an exemption from all law.

Recommended Governance Practices

Basic documentation
Internal ownership
Basic monitoring
Incident response

Classification Decision Tree

Classification Table

For each AI system from the inventory:

AI ID	Name	Annex III?	Art. 6(1) route?	Art. 50 cases?	Classification	Note
AI-001		[ ] Yes [ ] No	[ ] Yes [ ] No	[ ] Yes [ ] No
AI-002		[ ] Yes [ ] No	[ ] Yes [ ] No	[ ] Yes [ ] No

Next Steps

After classification:

Applicable prohibited practice -> Stop the prohibited conduct and consult Legal
High-Risk -> Applicable duties and DPIA screening
Transparency -> Implement applicable Article 50 measures, including alongside high-risk duties
Minimal -> Recommended basic documentation, after assessing applicable duties

Resources

Primary sources: AI Act — informative consolidated version 27 July 2026; Regulation (EU) 2026/1744 — published 24 July 2026; GDPR — published version 4 May 2016.